📄 Sentence Reduction for Automatic Text Summarization

Summary

Most extractive summarizers copy whole sentences, while professional abstractors usually trim them. This paper addresses sentence reduction: deciding which parts of an extracted sentence can be dropped without losing its main point. These parts can be words, prepositional phrases, gerunds, to-infinitives or clauses. The system takes the extracted sentence and its source document, parses the sentence with the IBM ESG parser, and annotates each parse-tree node with three kinds of evidence:

• Grammatical necessity, from thematic-role rules and a combined lexicon of verb subcategorizations.
• Local-context relevance, from weighted counts of lexical links such as repetition, morphology and WordNet relations.
• Corpus probabilities that a subtree is removed, partially reduced or left intact.

A final top-down pass combines these signals into a decision for each subtree. The training and test data are 500 sentences paired with professional reductions, mined automatically from Benton Foundation news abstracts. Humans shortened these sentences by 44.2% on average. The system's decisions agreed with the human ones 81.3% of the time, which suggests reduction can make extractive summaries much more concise and coherent.

Key points

• The paper builds on an analysis of human summaries that found six editing operations. Reduction is one of them, alongside combination, syntactic transformation, paraphrasing, generalization/specification and reordering.
• The pipeline has five steps: ESG parsing, grammar checking, context scoring, corpus evidence and a final decision pass.
• The lexicon merges COMLEX, Levin's English Verb Classes and Alternations, WordNet, and the WordNet-sense-tagged Brown Corpus. It covers subcategorization frames for over 5,000 verbs and is used to protect obligatory verb arguments.
• The corpus was built with the authors' decomposition program, which aligns phrases in human abstracts with the source articles. The articles are telecommunication news from the Benton Foundation's "Communications-related headlines" service.
• On average, humans shortened the 500 sentences by 44.2%, and the system matched 81.3% of human reduction decisions.

Interesting ideas

• I1. Harvesting training pairs from existing summaries: aligning human abstracts back to their sources yields original/reduced sentence pairs with no dedicated annotation effort. A similar alignment trick could build compression data for Ukrainian from news digests.
• I2. Obligatoriness relative to the parent node: each node is marked as required or optional only with respect to its parent, so the tree can be pruned locally and safely. This maps naturally onto dependency trees such as UD.
• I3. Head-conditioned deletion probabilities with three outcomes: the model estimates how likely a subtree is to be removed, partly reduced or kept, given the head word. This captures lexical habits, such as editors almost never touching the content clause after reporting verbs, which purely syntactic rules would miss.
• I4. Typed, weighted lexical links as a topicality signal: exact repetition counts more than a weaker WordNet relation like hypernymy, and phrase scores are summed up the tree. This is a cheap, interpretable salience measure that could feed other selection or ranking tasks.
• I5. Evaluating individual decisions rather than whole outputs: measuring agreement on each keep/drop choice against human edits gives a fine-grained score. That score is easier to diagnose than judging overall summary quality.